Physics-informed neural network for prediction of scour depth using natural frequency of monopiles

Abstract

Scour development around monopiles supporting offshore wind turbines (OWTs) poses a significant threat to the integrity of OWTs. An accurate and efficient scour depth prediction model is of great necessity to ensure the safety and reliability of offshore wind farm. This study proposes a physics-informed neural network (PINN) model for scour depth prediction, developed through inverse analysis of the natural frequency of OWT. The model incorporates physical laws into the data-driven framework through embedding the residual error of physical equations into the loss function. The physical equations are considered as the relationship between natural frequency and scour depth provided by the previous research. The grid search and cross validation techniques are used to select the hyperparameters, including the optimal number of hidden layers, neurons per layer and training epochs. The accuracy of the proposed model is rigorously validated under four distinct sand conditions: dense sand, medium compact sand, loose sand and very loose sand. Comparative analysis demonstrates that the PINN model achieves lower root mean square error (RMSE) than physical equations across all conditions, highlighting its superior accuracy and extrapolation capability. Furthermore, validation against field-monitored scour depth and natural frequency data confirms the accuracy and applicability of the PINN model.